In new research, Saharsh Agarwal and Ananya Sen study how Google AI Overviews reduce traffic to content publishers, their impact on consumer experience, and the implications of these findings for platform regulation and copyright and competition law.
Online search has rested on a remarkably stable symbiotic relationship over the past two decades. Publishers and other websites produce content, and search engines help users discover it. Search engines benefit from access to a large, continually updated ecosystem of third-party content, while publishers benefit from the traffic search engines drive to their websites. That downstream traffic to publisher websites can generate advertising revenue, subscriptions, first-party data, brand recognition, and longer-term relationships with consumers.
Generative artificial intelligence has the potential to fundamentally reshape this relationship. When Google introduced AI-generated summaries (AI Overviews or AIOs) into Search, the promise to key stakeholders in the online information ecosystem seemed straightforward. Instead of forcing users to sift through multiple websites, AI could synthesize the answer instantly. Everyone would win. Users would save time, searches would be more efficient, and publishers would receive traffic from citations embedded in summaries.
However, the question of whether AI summaries improve discovery or mainly reduce visits to the websites that produce the underlying information remains open. Further, if users stop clicking through, are the gains in convenience large enough to justify the resulting shift in value across the digital ecosystem? Answers to these questions have implications for platform design and public policy.
Our recent research deploys a field experiment to provide the first causal evidence to begin answering these questions. Using a custom Chrome extension, we randomly assigned users to either standard Google Search, in which AIOs appeared when triggered by a search prompt, or a version in which AIOs were removed in real time. We recruited over 1,000 users who installed this extension in early 2026 and observed their behavior for two weeks. Our baseline result shows the appearance of an AIO reduced users’ organic clicks to third-party sites by 39.8 percent and increased searches where the user clicked on no links at all by 34.5 percent. Additionally, we found no corresponding change in clicks on sponsored links or in search frequency.
Our browser extension also allowed us to measure, conditional on a click, bounce rates and time spent on the downstream or third-party publisher website as measures of click quality. Our data show no difference across these click quality measures between the treatment and control groups, seemingly contradicting Google’s public claims that AIOs produced higher quality traffic for websites that are clicked through on Google. Heterogeneity in the treatment effect suggests the impact on clicks to third-party sites is driven by instances where the AIO appears at the top of the search engine results page, in line with prior evidence that high-ranked content captures large amounts of user attention. We also fielded a post-experiment survey to quantify users’ stated search experience during the experimental period. We found no measurable improvement in users’ perceived search quality or ease of finding information. Exploratory evidence from Google’s conversational AI Mode suggests that the reduction in traffic to third-party sites could become even greater as search becomes more AI-mediated.
Our experiment does establish an economically important mechanism that has implications for current legal and policy debates: AI-generated answers can substitute for visits to external sources that help sustain the information ecosystem without measurable improvements in user experience or the quality of engagement. While the purpose of this study was not to provide the optimal legal remedy, we consider how our findings inform current policy and legal debates involving search platforms and publishers.
Disputes between platforms and publishers are nothing new. Earlier conflicts over news aggregation, snippets, linking, and publisher compensation raised similar questions about whether platforms should pay for content and whether publishers could meaningfully opt out. Generative AI has heighted the nature of the conflict.
One emerging policy response to Google AI Overviews, like the one proposed by the United Kingdom Competition and Markets Authority, is to give publishers greater control over whether their content can be used in AI Overviews. Such measures are valuable since publishers should have meaningful information and input about how their content is used. Search ranking and other platform algorithms have historically been opaque, and greater transparency and control can improve accountability.
But an opt-out, by itself, may not address the underlying economic problem identified by our experiment: the presence of the AI-generated answer substantially reduces the probability that users visit external websites at all. Moreover, in our data, citations within AIOs generated relatively little referral traffic (about 8% of all clicks). Whether the publisher opts in or out, it is unlikely to receive any traffic with the presence of AIOs.
Additionally, it is unclear how opting out of AIOs will affect publishers’ prominence in organic search rankings. If, over the longer term, there is a reinforcing loop between presence in AIOs and organic search rankings, publishers may face a difficult participation dilemma. This is particularly important because traffic has value beyond the revenue generated by a single page view. Website visits can help publishers build direct relationships with users, convert occasional users into subscribers, develop brand recognition, and learn about audience preferences. A remedy that focuses only on whether a publisher is cited, even through a pay-per-citation model, may therefore miss a substantial part of the economic value being redistributed to platforms.
Our findings map onto competition policy and conversations about platforms self-preferencing their products and services. When the same firm controls an important channel of discovery and decides how prominently its own AI-generated answers appear relative to external sources, questions about platform incentives cannot be avoided. Because AIOs are a Google product derived from external sources and occupy prominent search-page real estate, they may raise self-preferencing concerns among regulators (e.g., as in the EU). Indeed, our results show that platform design can materially shape the distribution of traffic and economic value. The reduction in external clicks was concentrated in cases where AIOs occupied the most prominent position above traditional organic results. Our study does not address several essential elements of recent legal cases, such as market definition and exclusionary conduct, nor does it establish an antitrust violation.
AIOs also raise questions about copyright. It is clear that publishers incur the costs of producing information, while Google’s AI systems utilize external content to generate answers. The experiment does not establish whether an output is substantially similar to a copyrighted work or whether particular uses are legally fair. Those legal questions require substantially more analysis than our paper can provide. Nevertheless, our findings may provide a useful starting point. For example, copyright law has long been concerned not only with copying, but also with the extent to which a new use substitutes for the original and affects the copyright holder’s market. If courts or policymakers consider such market effects, the economic mechanism we document may be relevant.
Ultimately, the appropriate remedy should focus on the structural issues that will allow a healthy online information ecosystem to flourish. Opt-outs and transparency address publisher consent and control but are insufficient if a central concern is lost traffic. Platform design remedies aimed at preserving traffic could include making source links substantially more prominent and experimenting with interface designs that preserve meaningful pathways to original content. Because our results indicate that the effects are strongest when AIOs receive the most prominent placement, remedy design should pay attention not only to whether links exist but also to whether users are realistically likely to engage with them. Given the relatively low share of clicks generated by citations within AI Overviews, a pay-per-citation model could serve as an intermediate remedy, but it would be insufficient if click-through traffic is necessary for publishers to build direct user relationships
Compensation and licensing provide another possible response that may help preserve incentives to invest in high-quality content. Licensing payments may offset some lost advertising revenue but not the benefits of direct user relationships, subscription conversion, audience data, and brand building. Even if these elements can be priced in, licensing could leave downstream publishers highly financially dependent on Google. Indeed, such dynamics might explain why websites like Reddit have been hesitant to renew their licensing deals with Google. Moreover, large publishers may be able to negotiate licensing agreements, while smaller publishers lack sufficient bargaining power. This might concentrate both compensation and visibility among already powerful content providers. The opposite problem is also possible. If high-quality publishers broadly withhold their content from AI systems, the quality and usefulness of AI-generated answers may deteriorate, as has been conjectured in recent research. Policy must therefore account for both sides of the ecosystem: preserving incentives to create valuable information while allowing socially useful innovation in search.
That is the central implication of our experiment: industry stakeholders and regulators should work to ensure that innovation in information access does not systematically undermine the incentives to produce the information on which those innovations depend. Existing legal frameworks may ultimately address parts of this problem through copyright, competition law, or emerging platform regulation. But whatever doctrinal route policymakers choose, remedies should be evaluated against the economic mechanism actually occurring in the market, which our study focuses on.
Authors’ Disclosures: The authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
